If you can reduce complex matrix multiplications into simpler terms, then you may be able to focus the training based on those constraints to increase performance/efficiency.